How to measure AI search visibility: three numbers that never agree

Measure AI search visibility three ways: citations for fixed questions, AI referral visits, and Bing and Google reports. Why they disagree, with our numbers.

· Updated 8 min read

You measure AI search visibility with three separate numbers: how often AI answers to a fixed set of questions cite you, how many visits AI assistants send according to your analytics, and what first-party reports from Bing and Google count. They measure different things and they rarely agree, so no one of them is the answer.

What can you actually measure?

MethodWhat it countsWhat it misses
Citation scan of a fixed question setThe share of answers that cite or mention you, per engineEvery question you did not think to ask
AI referral visits in analyticsPeople who clicked through from an assistantVisits whose referrer was stripped, and citations nobody clicked
First-party reports (Bing, Search Console)Citations or impressions the vendor counted itselfEvery other engine, the question behind each one, and clicks

A citation is a link to your page given as a source. A mention is your name in the answer with no link. Count them apart: a mention builds recognition, and only a citation can send a visit.

Why don't the three numbers agree?

Here are our own sites, measured three ways over overlapping periods.

Citation scan. In the weekly scan of 14–15 November 2026, across ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity and Claude, citelity.com was cited in 0 of 144 answers to 24 tracked questions. website.show was cited in 0 of 78 answers to 13 questions.

Analytics. In the same 28 days, Google Analytics recorded 67 visits to website.show from ChatGPT, almost all to one page.

First-party report. Bing Webmaster Tools reported 261 Copilot citations for citelity.com between 1 October and 12 November 2026.

So one site has zero citations and 67 visits from the same assistant, and the other has zero citations and 261. None of these numbers is wrong. The scan asked the questions we chose, and ChatGPT was sending people to website.show for questions we were not tracking. Copilot is not one of the six engines in the scan at all, and its citations came from searches we would not have guessed. Each method sees one surface.

The useful reading is in the gaps. Visits without tracked citations mean your question set is missing the questions you are already cited for; the page those visits land on tells you where to look. Citations in a first-party report with no visits mean you are being used as a source and not clicked.

How do you run a citation scan?

  1. Write the questions as people ask them. Full questions, not keywords: the things a buyer or reader would type into an assistant, including comparisons and "best X for Y". A few dozen is enough to see a pattern; five is not.
  2. Keep the set fixed. A share of answers only compares with last week's if the questions are the same. When you add or drop questions, mark the scan, and do not read the change as movement.
  3. Count per engine. The engines draw on different indexes and cite different sites. A single blended score hides the engine where you are absent.
  4. Repeat on a schedule. An assistant can give a different answer to the same question minutes later, so one answer is one sample. A weekly scan of the whole set smooths that out; a single screenshot proves only that a citation is possible.
  5. Record who was cited instead. The sites that appear where you do not are the list to study. That is worth more than the percentage.

Our methodology page describes how citelity runs these scans and labels where each engine's data comes from.

How do you read AI referral traffic?

In Google Analytics, filter sessions by source for the assistants' domains, such as chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com, and look at the landing pages.

Read the total as a floor. Assistants and in-app browsers often pass no referrer, and those visits are recorded as direct traffic. The real number is higher than the one you see, and you cannot know by how much.

The landing pages matter more than the total. On website.show nearly all of the 67 visits went to one page, which told us more about what ChatGPT uses the site for than the scan did.

What do first-party reports add?

They are the only numbers counted at the source. Bing Webmaster Tools has an AI Performance report with citations per day, per page and per grounding query, covering Copilot and Bing's AI summaries; we describe it in how to get cited by Microsoft Copilot. Its limits are that it covers one family of products, counts citations with no clicks, and uses Bing's own definition of a citation.

Google added its own in June 2026. Search Console now has a generative AI performance report that counts how often links to your pages were shown in AI Overviews and AI Mode, with the pages and breakdowns by country, device and date (Search Engine Journal). It is the first number for Google's AI surfaces that comes from Google. Read it with three limits in mind: it shows impressions and no clicks, it does not separate AI Overviews from AI Mode, and the same impressions are already inside the totals of the ordinary performance report, so the two must not be added together (Scroll).

Neither report tells you which question you were shown for. That is what the scan is for.

What should you not trust?

AI prompt volume. No assistant publishes how often a question is asked. A tool that shows "prompt volume" is estimating it from something else, usually search volume or a panel. Search volume for the underlying query is a real number and useful for ordering your questions; it is not a count of AI prompts.

A single visibility score. A blended number across engines and questions moves when you change the question set, and it hides which engine and which question changed.

One answer. A screenshot of an assistant citing you, or not citing you, is one sample.

How do I measure AI search visibility?
With three numbers side by side: the share of AI answers to a fixed question set that cite you, per engine; visits from AI assistants in your analytics; and first-party reports from Bing Webmaster Tools and Google Search Console. They measure different things, so read each for what it can see.
Why do my AI visibility numbers disagree with each other?
Each method watches a different surface. A scan sees only the questions you chose, analytics sees only clicks that kept their referrer, and a first-party report covers one vendor. On our own sites a scan found 0 citations while analytics showed 67 ChatGPT visits and Bing reported 261 Copilot citations.
How often should I check AI citations?
On a fixed schedule with the same questions, weekly for most sites. One answer is a single sample, because an assistant can answer the same question differently minutes later.
Is AI referral traffic in Google Analytics accurate?
It is a floor. Visits where the assistant or an in-app browser passed no referrer are recorded as direct traffic, so the real number is higher than the one shown.
Can I see AI prompt volume for a keyword?
No assistant publishes it. Tools that show prompt volume estimate it from search volume or panels. Use search volume to order your questions, not as a count of AI prompts.
What is the difference between a mention and a citation?
A citation is a link to your page given as a source. A mention is your name in the answer without a link. Only a citation can send a visit, so count them separately.

Sources